Echo: A multi-agent AI system for patient-centered pharmacovigilance

Megha Srivastava

Keywords: pharmacovigilance, multi-agent, biomedicine, social media

TL;DR: Echo is a multi-agent AI system that mines Reddit posts to identify drug-symptom associations missing from official databases. From only 187 posts, it extracted 640 associations, including novel (drug, ADR) relatif

Abstract:

Online health communities provide patients with spaces to share experiences, find
support, and voice concerns that may go unacknowledged in clinical encounters.
These narratives often include early reports of adverse drug reactions (ADRs), yet
remain largely absent from formal pharmacovigilance. We present Echo, a multi-
agent AI system that transforms patient narratives from Reddit into structured drug
safety intelligence. Echo deploys four specialized language model agents in concert:
an Explorer mining social media forums, an Analyzer quantifying associations
through temporal, confidence, and community metrics, a Verifier identifying novel
signals absent from FDA databases, and a Proposer generating testable hypotheses
from biomedical literature. As a proof-of-concept, we show that from less than 200
Reddit posts, Echo was able to discover 640 drug-symptom associations, including
several absent from official databases, such as pembrolizumab-induced daytime
somnolence. We further show in retrospective case studies that Echo might have
detected emerging toxicities, such as checkpoint inhibitor pneumonitis, before
regulatory recognition. Beyond signal detection, Echo also identifies confounding
factors and proposes testable hypotheses. Finally, we build an interactive interface
to help explore associations, examine patient quotes, and access AI-generated
insights. Overall, Echo leverages language models to surface patient-reported
signals that may complement regulatory surveillance.

Submission Number: 287